AI Dev 25 X NYC | Jacky Liang: Why Agents Can't Find the Right Docs (And How Postgres Fixes It)

Why It Matters

Hybrid search inside PostgreSQL eliminates costly multi‑system pipelines, delivering more accurate AI‑driven documentation retrieval while slashing operational overhead and cloud spend.

Summary

Jacky Liang, a developer advocate at Tiger Data (TimescaleDB), opened the session by highlighting a persistent problem in AI‑augmented search: pure vector‑only retrieval often returns semantically similar but factually incorrect documentation, especially when version numbers or API signatures change. He contrasted vector search—relying on high‑dimensional embeddings that capture meaning—with traditional keyword search, which excels at exact term matching but fails on synonyms and contextual nuance. The core argument is that neither approach alone satisfies the precision demands of LLM‑driven agents that need to fetch the right snippet from evolving technical docs.

Liang walked through concrete failure modes: a docs chatbot might suggest deprecated PostgreSQL 16 parameters when the user is on version 17, a coding assistant could propose syntax that no longer exists, and a troubleshooting bot may recommend fixes for bugs that have been resolved. He illustrated how keyword search (BM25) and vector search can be fused using Reciprocal Rank Fusion (RRF) to produce a hybrid pipeline that preserves exact matches while still leveraging semantic similarity. The hybrid model returns the correct “Postgres 17 Max Connections Config” result, whereas a vector‑only query would surface irrelevant PostgreSQL 16 or MySQL entries.

The technical deep‑dive then shifted to implementation. Liang critiqued the typical architecture that stitches together a relational store, a separate vector database (e.g., Pinecone, Qdrant), and an Elasticsearch‑style keyword engine, noting the operational overhead of ETL pipelines, eventual consistency glitches, and cost escalation. He advocated for using PostgreSQL itself as the unified platform, leveraging its built‑in full‑text search (TSVector/TSQuery) and modern BM25‑style ranking via the new PG Tech Search plugin. This plugin adds a modern ranked keyword engine to Postgres, eliminating the need for external services while still supporting hybrid search through RRF.

The takeaway for enterprises is clear: by adopting hybrid search directly within PostgreSQL, teams can cut infrastructure complexity, reduce latency, and improve retrieval accuracy for AI agents. The newly released PG Tech Search plugin positions TimescaleDB as a one‑stop solution for RAG‑style applications, promising tighter consistency, lower total cost of ownership, and easier scaling for documentation‑heavy workloads.

Original Description

At AI Dev 25 x NYC, Jacky Liang, Developer Advocate at TigerData, showcased hy AI agents struggle to find the correct documentation and how to fix this using PostgreSQL.

Most production RAG systems fail because vector search alone hallucinates or misses exact term matches in legal clauses, technical specs, and domain-specific content.

This technical talk shows how to implement hybrid retrieval directly in Postgres by combining BM25 statistical ranking with pgvector semantic search.

Attendees will learn how to build a working RAG pipeline that queries knowledge bases with both search phrases and embedding vectors, returning ranked results weighted by both keyword relevance and semantic similarity—eliminating the need for external search infrastructure like Elasticsearch.

----------

Join us at AI Dev 26 x San Francisco! Tickets: https://ai-dev.deeplearning.ai